Measuring AI search: GA4 AI referrals, Search Console and a monthly citation panel

You can measure AI search with four instruments that cost nothing but time: GA4's built-in "AI Assistant" channel plus a custom channel group for the referrers it misses; Search Console, where Google reports AI Overviews and AI Mode inside the "Web" search type rather than separately; a monthly citation panel of 20 prompts run by hand across four engines; and your CDN logs, which show whether AI search crawlers are fetching the pages you want cited. None is complete on its own. Together they answer the three questions that matter: are we being cited, are we being visited, and are we being crawled.

Key takeaways

  • GA4's default channel grouping includes "AI Assistant", defined by Google as "the channel by which users arrive at your site from sources like ChatGPT"; it matches a medium of ai-assistant or a referrer on Google's list of assistants. It is not where Google's own AI Overviews and AI Mode clicks land.
  • Google says AI Overviews and AI Mode are "reported on in the Performance report, within the 'Web' search type"; there is no separate report. Bing, by contrast, launched an AI Performance report in public preview in February 2026 with citation counts and the queries that grounded them.
  • Visits are a small, lagging signal: Cloudflare measured about 1,091 pages crawled per visit referred for OpenAI and about 38,066 for Anthropic in July 2025. Citations and crawls lead; count them.
  • A 20-prompt panel run on the same day each month is the only direct measure of whether you are cited. It takes about an hour.

Why most businesses cannot answer the question

Four blind spots, in our experience of auditing analytics setups:

  1. Grouping. Without a channel set up for assistants, their referrals sit inside Referral and Direct next to everything else.
  2. Google's AI features look like Google. A click from an AI Overview or AI Mode arrives from google.com like any other search click, so it lands in Organic Search with no marker.
  3. Lost referrers. Links copied out of an assistant, opened in an app's in-app browser or stripped of parameters arrive as Direct. ChatGPT appends utm_source=chatgpt.com to many of its citation links, but secondary reports agree it is not on every link, so what you count is a floor.
  4. Nobody records citations. The engines do not tell you when they quote you. Only a panel does.

The measurement below is the fourth step of our GEO guide and is what we run inside the AI search visibility service.

Instrument 1: GA4, in six steps

  1. Look at what is already there. Reports, Acquisition, Traffic acquisition; set the dimension to Session default channel group and find "AI Assistant". Google's definition: the medium exactly matches ai-assistant, or the medium is ai-assistant with campaign (ai-assistant) when the referrer matches its list of AI assistants (Google names ChatGPT, Gemini and Copilot among the sources).
  2. Create a custom channel group. Admin, Data display, Channel groups, Create new channel group. A standard property allows two custom groups of up to 50 channels, and Google states that "custom channel groups can be applied to your reports retroactively", so history is covered from the day you create it.
  3. Add an "AI referrals" channel at the top of the order (traffic is assigned to the first channel it matches). Condition: Source matches the regular expression chatgpt\.com|chat\.openai\.com|perplexity\.ai|copilot\.microsoft\.com|gemini\.google\.com|claude\.ai. Google's own documentation uses a regex on source for its AI-assistant example. Extend the list as you see new sources in the Referral report.
  4. Add a second channel whose condition is Default channel group exactly matches "AI Assistant", so Google's own classification is kept alongside yours.
  5. Build a landing-page exploration. Explore, Free form; rows: Landing page + query string; columns: Session custom channel group (your new group); metrics: Sessions, Engaged sessions, Key events. Save it; this is the page you open monthly.
  6. Mark your key events (generate_lead, booked calls, phone and WhatsApp clicks) so the exploration shows whether AI referrals convert, not only whether they arrive. A property with no key events cannot answer the only question the owner will ask.

Instrument 2: Search Console and Bing Webmaster Tools

Search Console. Google's AI features page says AI Overviews and AI Mode traffic is "reported on in the Performance report, within the 'Web' search type", and that no separate report exists. The Performance report's search types are "Web: text-based", "Web: multimodal", Image, Video and News. So the method is comparison, not isolation: filter by page to the URLs you rewrote for AI visibility, compare clicks and impressions against the previous period, and export the page table monthly so you keep your own history.

Bing Webmaster Tools. On 10 February 2026 Microsoft announced AI Performance, in public preview, showing "how publisher content appears across Microsoft Copilot", AI-generated summaries in Bing and selected partner integrations. Its metrics are Total Citations, Average Cited Pages, Grounding Queries (the phrases the AI used when retrieving your content), page-level citation counts and a trend line. Microsoft's caveats are worth repeating: citation counts "don't show where content appears within an answer", and Grounding Queries are a sample. It is still the only first-party citation report from a major engine as of December 2026, and it is free.

Instrument 3: the monthly citation panel

This is the direct measurement. It is manual, and at 20 prompts that is a feature: an hour a month, no scraping, no terms-of-service questions.

Design. Twenty prompts your buyers would type: ten informational ("how do I migrate a WordPress site without losing rankings") and ten commercial ("which agency should I use for a Shopify store in Ontario", "best eCommerce developer near Niagara Falls"). Four engines: ChatGPT with search on, Perplexity, Gemini and Google AI Mode; add Copilot if Bing matters to you. Same prompts, same day each month, in a clean browser profile, from the location your buyers are in. Record one row per prompt per engine.

Column What to record Why
Date, engine, location Exactly as run Answers vary by day and place; you are tracking a series, not a truth
Prompt Verbatim Keep the set stable for at least six months
Cited Yes or no The headline metric: citation share = cells cited / total cells
URL cited The exact page Tells you which page earns citations, which rarely matches the page you expected
Position Order among the answer's citations Being first is worth more than being eighth
Competitors cited Domains Who the engine prefers for this question
Answer summary One line Spots factual errors about you and the claims competitors are credited with

Scoring. Report citation share overall and per engine, the list of cited URLs, and the top ten competitor domains. In month one you are establishing a baseline; the first useful comparison is month three. Any factual error about your business in an answer goes straight into a content fix, because engines quote what they can find.

Instrument 4: crawler logs

An engine cannot cite a page its search crawler has not fetched. On CloudFront, enable standard logging (v2) to S3 (AWS states there are "no additional charges for log delivery to Amazon S3"; you pay for storage), create the Athena table AWS publishes for CloudFront standard logs, and count requests per bot and per page each month. The user-agent strings to look for are OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot and bingbot for search, and ChatGPT-User, Perplexity-User and Claude-User for visits made on a user's behalf. The full query and the bot table are in our AI crawlers and robots.txt guide. Our target for every new URL we publish: fetched by at least OAI-SearchBot and PerplexityBot within 30 days. If a page misses it, the problem is access or discovery (robots.txt, CDN rules, internal links, sitemap), not content.

The one-page monthly report

Everything above fits on one page, and one page is what gets read:

  • AI referral sessions, engaged sessions and key events by landing page (GA4 exploration), with the "AI Assistant" default channel shown for comparison.
  • Search Console clicks and impressions on the target pages, this month against last.
  • Bing AI Performance citations and top grounding queries.
  • Citation panel: share overall and by engine, cited URLs, competitor domains, errors found.
  • Crawl coverage: the percentage of target URLs fetched by each search crawler in the last 30 days.
  • Five actions for next month, each tied to one of the numbers above.

Expect the first baseline to be small. For most service businesses the "AI Assistant" channel shows a handful of sessions a year, often a number nobody was looking at; record it anyway, because month one is the line every later report is measured against.

What to expect

Set expectations with the numbers that exist. Cloudflare's July 2025 measurement of crawls per referred visit (Google 5.4, Microsoft 40.7, Perplexity 194.8, OpenAI 1,091.4, Anthropic 38,065.7) says that even a well-cited site will see few assistant referrals relative to how often it is read by machines, and that 79% of AI crawler requests were for training rather than search. The visits that do arrive tend to be from people who have already read a comparison, which is why key events per session, not sessions, is the figure to watch. Citations move first, after the search crawlers have refetched a rewritten page; referrals follow; conversions are the last and smallest series. A programme that only looks at GA4 will conclude nothing is happening for months while the panel shows it is.

FAQ

Why can't I see Google AI Overviews traffic in GA4?

Because Google does not mark it. Clicks from AI Overviews and AI Mode arrive from google.com like any other search click and are assigned to Organic Search. Google reports those features only inside Search Console's Performance report under the "Web" search type, with no separate AI report, so compare page-level clicks over time rather than looking for an AI line in GA4.

Is GA4's "AI Assistant" channel enough on its own?

No. It is a useful default (Google defines it as traffic "from sources like ChatGPT" matched by an ai-assistant medium or a referrer on Google's list), but it misses referrers not on that list, visits whose referrer was stripped, and Google's own AI features. Add a custom channel group with your own source regex and a landing-page exploration; custom groups apply retroactively.

Why does some ChatGPT traffic show up as Direct?

Because the referrer is lost. Links copied out of the app, opened in in-app browsers or stripped of parameters carry no source. ChatGPT adds utm_source=chatgpt.com to many citation links, which GA4 then classifies correctly, but secondary reports agree it is not on every link. Treat the AI referral count as a floor.

How many prompts does a citation panel need?

Twenty is enough to see movement at the level of a small or mid-size business: ten informational and ten commercial prompts, across four engines, gives 80 observations a month. Keep the set fixed for at least six months; change prompts and you reset the series. More prompts add precision but also hours; automate only after checking each engine's terms.

Can I automate the citation panel?

Tools exist that query the engines on a schedule, and some engines offer APIs with search. Check the terms of service of each engine before scraping its consumer interface, budget for API costs, and keep a manual run each quarter to check that the automated results match what a person sees in the same location.


Impinge Solutions sets up the GA4 channel group, the Search Console comparison and the monthly citation panel as part of its AI search visibility service. Ask for the measurement setup and the first baseline report.

Sources

Are you ready to partner with us?

Our good-natured team is waiting to hear your business requirements.

Book a free project call

Stay informed with our newsletter